anton/cmpo/README.md

CMPO Integration Module

Overview

The Cellular Microscopy Phenotype Ontology (CMPO) integration module is a core component of Anton that provides semantic mapping between natural language descriptions and standardized scientific terminology. This module enables Anton to translate VLM-generated insights into scientifically compliant, searchable, and interoperable phenotype classifications.

Problem Statement

Modern microscopy analysis faces a critical challenge: bridging the semantic gap between AI-generated natural language descriptions and standardized scientific terminology. While VLMs can provide expert-level biological insights ("cells arrested in metaphase with condensed chromosomes"), these descriptions need to be mapped to formal ontology terms for:

• Scientific standardization: Ensuring consistent terminology across studies
• Data interoperability: Enabling cross-dataset comparisons and meta-analyses
• Knowledge integration: Connecting observations to broader biological knowledge graphs
• Reproducible research: Providing precise, unambiguous phenotype classifications

Conceptual Framework

1. Multi-Level Hierarchical Mapping

CMPO is organized in a hierarchical structure with multiple branches:

CMPO Root
├── biological_process (GO terms)
├── cellular_phenotype (398 terms)
│   ├── cell_population_phenotype (73)
│   ├── cell_process_phenotype (157)
│   │   ├── cell_cycle_phenotype (46)
│   │   │   ├── cell_cycle_arrested_phenotype (6)
│   │   │   │   ├── G2_arrested_phenotype
│   │   │   │   ├── M_phase_arrested_phenotype
│   │   │   │   └── metaphase_arrested_phenotype
│   │   │   └── mitotic_process_phenotype (37)
│   │   └── cell_death_phenotype (1)
│   └── cellular_component_phenotype (186)
├── molecular_entity (CHEBI terms)
├── molecular_function (GO terms)
└── quality (PATO terms)

2. Research Context-Aware Subgraph Navigation

Key Insight: Researchers often have specific analytical intentions that determine which CMPO subgraphs are most relevant.

Context Types:

• Process-focused: Studying cell division, apoptosis, migration → cell_process_phenotype subgraph
• Component-focused: Analyzing organelles, structures → cellular_component_phenotype subgraph
• Multi-intent: Cell cycle AND mitochondrial analysis → Multiple overlapping subgraphs
• Population-level: Colony behavior, density effects → cell_population_phenotype subgraph

3. Two-Strategy VLM Mapping Approach

Strategy 1: Description → CMPO Mapping

VLM Analysis: "Cells show metaphase arrest with hyperconnected chromosomes"
    ↓
Semantic Parsing: Extract ['metaphase', 'arrest', 'chromosomes', 'condensed']
    ↓
CMPO Mapping: → CMPO:0000XXX "metaphase arrested phenotype"

Strategy 2: CMPO-Guided Evidence Detection

Research Context: "Studying cell cycle defects"
    ↓
Subgraph Selection: Focus on cell_cycle_phenotype branch
    ↓
VLM Query: "Do you see evidence of: metaphase arrest, anaphase defects, etc.?"
    ↓  
Targeted Classification: Direct mapping to specific terms

Technical Implementation

Semantic Mapping Pipeline

1. Ontology Loading: Parse full CMPO .obo file with rich semantic relations
2. Multi-Modal Matching:
   • Direct matching: Term names and synonyms
   • Semantic matching: Logical definitions and cross-ontology references
   • Contextual matching: Hierarchical subgraph relevance
3. Confidence Scoring: Weighted combination of multiple evidence sources
4. Hierarchy Navigation: Maintain relationships for downstream analysis

Rich Ontological Information

Each CMPO term contains:

{
    "CMPO:0001234": {
        "name": "metaphase arrested phenotype",
        "description": "A phenotype in which cells are arrested in metaphase",
        "synonyms": ["metaphase arrest", "M-phase block"],
        "subclass_of": ["cell_cycle_arrested_phenotype", "mitotic_phenotype"],
        "equivalent_to": "has_part(arrested and characteristic_of(mitotic_metaphase))",
        "xrefs": ["GO:0000819"],  # Cross-ontology links
        "subset": ["cmpo_core"]
    }
}

Two-Stage Mapping Pipeline

async def map_to_cmpo_enhanced(description, cmpo_ontology, vlm_interface, context=None):
    # Stage 1: Ontology-Aware Candidate Generation
    candidates = ontology_aware_mapping(description, cmpo_ontology, context)
    
    # Stage 2: VLM Biological Reasoning & Pruning
    if len(candidates) > 1:
        validated_mappings = await vlm_biological_validation(description, candidates, vlm_interface)
        return validated_mappings
    else:
        return candidates

def ontology_aware_mapping(description, cmpo_ontology, context=None):
    # 1. Enhanced token extraction with exact matching priority
    exact_tokens = extract_exact_biological_matches(description)
    fuzzy_tokens = extract_fuzzy_biological_tokens(description)
    
    # 2. Hierarchical scoring
    for term_id, term_data in cmpo_ontology.ontology.items():
        score = 0
        
        # Exact token matches (highest weight)
        exact_score = calculate_exact_matches(exact_tokens, term_data) * 1.0
        
        # Hierarchical specificity (deeper = more specific = higher score)
        specificity_score = calculate_hierarchy_depth(term_id, cmpo_ontology) * 0.3
        
        # Ontological distance (closer = more related = higher score)
        distance_score = calculate_ontological_distance(term_id, context_terms) * 0.2
        
        # Fuzzy similarity (lowest weight)
        fuzzy_score = calculate_fuzzy_similarity(fuzzy_tokens, term_data) * 0.1
        
        total_score = exact_score + specificity_score + distance_score + fuzzy_score
    
    return ranked_candidates

async def vlm_biological_validation(description, candidates, vlm_interface):
    validation_prompt = f"""
    Original biological description: "{description}"
    
    Candidate CMPO term mappings:
    {format_candidates_for_review(candidates)}
    
    Task: Evaluate biological plausibility and ranking of these mappings.
    
    Consider:
    - Biological consistency and logical compatibility
    - Temporal/spatial relationships in biological processes
    - Phenotypic co-occurrence patterns
    - Mechanistic plausibility
    - Specificity vs generality trade-offs
    
    Provide:
    1. Biologically valid mappings (with confidence 0-1)
    2. Brief scientific reasoning for each acceptance/rejection
    3. Final ranked list
    
    Focus on biological accuracy over textual similarity.
    """
    
    reasoning_result = await vlm_interface.reason_about_mappings(validation_prompt)
    return parse_and_apply_biological_reasoning(candidates, reasoning_result)

Usage Examples

Basic Mapping

from anton.cmpo import CMPOOntology, map_to_cmpo

cmpo = CMPOOntology()
results = map_to_cmpo("cells arrested in metaphase with condensed chromosomes", cmpo)

# Output:
# [
#   {
#     "CMPO_ID": "CMPO:0001234",
#     "term_name": "metaphase arrested phenotype", 
#     "confidence": 0.92,
#     "supporting_evidence": "Direct match: metaphase; Semantic: arrested + mitotic",
#     "hierarchy_path": ["metaphase arrested phenotype", "cell cycle arrested phenotype", "cell cycle phenotype"]
#   }
# ]

Context-Aware Mapping

# Research studying apoptosis
results = map_to_cmpo("fragmented nuclei with membrane blebbing", cmpo, context="apoptosis")
# → Higher confidence for apoptotic_cell_phenotype terms

# Research studying cell division  
results = map_to_cmpo("abnormal spindle formation", cmpo, context="cell_cycle")
# → Higher confidence for mitotic_process_phenotype terms

Integration with Anton Pipeline

# Within QualitativeAnalyzer
population_insights = await vlm.analyze_population(image)
cmpo_mappings = map_to_cmpo(
    description=population_insights['description'],
    cmpo_ontology=self.cmpo_mapper,
    context=self.research_context
)

Validation and Quality Assurance

Confidence Thresholds

• High confidence (>0.8): Direct term matches with strong semantic support
• Medium confidence (0.5-0.8): Semantic matches with contextual support
• Low confidence (0.3-0.5): Weak matches requiring human review
• Below threshold (<0.3): Excluded from results

Evidence Tracking

Each mapping includes:

• Supporting evidence: Specific text that triggered the match
• Mapping type: Direct, semantic, or contextual
• Hierarchy path: Full taxonomic classification
• Cross-references: Links to related GO/PATO terms

Future Enhancements

1. Machine Learning Integration

• Embedding-based similarity: Use biological language models (BioBERT, etc.)
• Context learning: Train models on researcher annotation patterns
• Active learning: Improve mappings based on user feedback

2. Advanced Semantic Reasoning

• Logical inference: Use formal ontology reasoning for complex mappings
• Negation handling: Detect and properly handle negative evidence
• Uncertainty quantification: Bayesian confidence estimates

3. Multi-Ontology Integration

• Cross-ontology alignment: Map to GO, PATO, CHEBI simultaneously
• Knowledge graph construction: Build comprehensive phenotype knowledge graphs
• Standardized interfaces: FAIR data principles compliance

4. Dynamic Ontology Updates

• Version management: Handle CMPO ontology updates gracefully
• Backward compatibility: Maintain mapping consistency across versions
• Community integration: Contribute mappings back to CMPO community

Research Applications

Enabled Use Cases

1. Large-scale phenotype screens: Standardized classification across thousands of images
2. Cross-study meta-analysis: Combine results from different research groups
3. Drug discovery: Map compound effects to standardized phenotype profiles
4. Disease research: Connect cellular phenotypes to pathological processes
5. Evolutionary studies: Compare phenotypes across species using common vocabulary

Scientific Impact

• Reproducibility: Eliminates ambiguity in phenotype descriptions
• Discoverability: Enables semantic search across phenotype databases
• Integration: Connects microscopy data to broader biological knowledge
• Collaboration: Provides common language for interdisciplinary research

---

Development Notes

Design Decisions

Why hierarchical subgraph mapping?

• CMPO contains >600 terms across diverse biological domains
• Research context dramatically improves mapping accuracy
• Enables both broad screening and focused deep analysis

Why two-strategy VLM approach?

• Strategy 1 (description→CMPO) handles unexpected discoveries
• Strategy 2 (CMPO-guided) ensures comprehensive coverage of known phenotypes
• Combination provides both discovery and validation capabilities

Why rich semantic relations?

• Simple keyword matching fails for scientific terminology
• Logical definitions enable precise semantic matching
• Cross-ontology links expand vocabulary and validation

Code Organization

• ontology.py: CMPO data loading, parsing, and management
• mapping.py: Core mapping algorithms and semantic analysis
• __init__.py: Module interface and public API
• README.md: Comprehensive documentation (this file)

Testing Strategy

• Unit tests for individual mapping functions
• Integration tests with full CMPO ontology
• Validation against expert-annotated datasets
• Performance benchmarks for large-scale analysis

---

This module represents a significant advancement in automated microscopy phenotype classification, bridging AI-generated insights with rigorous scientific standards.